Multi-modal data driven approaches for disaster damage assessment and prediction
Author(s)
Zhang, Danrong
Advisor(s)
Roy, Nimisha
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
As climate change accelerates, disasters pose an increasing threat to human lives and infrastructure. Disaster management is evolving from a reactive approach—addressing damage only after it occurs—to a proactive stage, where potential disasters are anticipated and preparations are made, and ultimately to a predictive stage, where data is used to forecast disaster impacts. While current disaster response remains largely reactive, with growing efforts towards proactive measures, this work addresses the gaps in reactive post-disaster damage assessments and advances the field towards proactive and predictive disaster management, with the goal of improving overall preparedness.
This thesis employs multi-modal data-driven methods to enhance both post-disaster damage assessment and pre-disaster damage prediction. By integrating Geographic Information Systems (GIS), data analytics, and machine learning with diverse data modalities, such as tabular data, social media imagery, satellite imagery, and nighttime light data, the study provides critical insights into disasters like hurricanes, tornadoes, earthquakes, and landslides. These approaches equip stakeholders with valuable information, reinforcing disaster preparedness and response strategies to mitigate future risks and enhance community resilience.
In the field of post-disaster damage assessment, this thesis addresses critical gaps by integrating alternative data sources and objective methodologies to assist prompt and accurate disaster response. Post-disaster damage assessments have largely relied on optical imagery, whether from satellites, drones, or social media, due to the intuitive nature of visual evidence in assessing damage. However, this focus on optical data can overlook other valuable sources of information. Technologies like Light Detection and Ranging (LIDAR) and Interferometric Synthetic Aperture Radar (InSAR), which capture 3D structural data and ground deformation, respectively, play critical roles in areas where optical imagery falls short. In the work "Black Marble Nighttime Light (NTL) Data for Disaster Damage Assessment," the application of NTL data was explored in assessing damage from hurricanes, tornadoes, and earthquakes. The findings revealed that NTL data is particularly effective in identifying hurricane-affected areas needing assistance, thereby enhancing the relief efforts. To address the subjectivity in traditional, human-interpreted damage degree classification, the work "From Pixels to Impact: Estimating Earthquake Damage Severity via Semantic Segmentation of Social Media Images" reframed damage assessment as a semantic segmentation problem. A pixel-level evaluation method with Segformer was developed, providing a more objective and standardized framework for consistent post-disaster reconnaissance.
In the field of pre-disaster damage prediction, this thesis focuses on leveraging advanced machine learning techniques to enhance disaster preparedness. In the work “Enhancing Landslide Susceptibility Mapping (LSM) Using a Positive-Unlabeled (PU) Machine Learning Approach”, instead of considering LSM as a binary classification problem, this work utilized PU learning, which treated areas with no historical landslides as unlabeled rather than negative instances. This approach improves the performance of susceptibility maps, aiding local governments in landslides preparedness. In ‘Predicting Hurricane-Induced Building Damage Using Multimodal Machine Learning: Insights from the StEER Dataset’ and ‘Predicting Tornado-Induced Building Damage: A Comparative Study of Tree-Based Models and Graph Neural Networks’, building damage data from the Structural Extreme Events Reconnaissance (StEER) dataset was explored. One study used multimodal machine learning to integrate pre-disaster data sources, while the other employed graph neural networks to model building interconnections. Both works provide building-level damage predictions, enhancing proactive disaster preparedness in hurricane and tornado scenarios.
This thesis also involves machine learning-aided sensor optimization. In ‘A Data-driven Approach to Optimize the Design Configuration of Multi-Sleeve Cone Penetrometer Probe Attachments’, multi-sleeve Cone Penetration Test (CPT) devices were optimized using a data-driven approach. Sensor configurations were refined to improve soil classification performance while minimizing the complexity, enhancing device efficiency and offering unique data for better pre- and post-disaster soil analysis.
Sponsor
Date
2024-11-19
Extent
Resource Type
Text
Resource Subtype
Dissertation (PhD)